CNT Hazard Probe v1: Synthetic Accelerator Hazard Memory

This record releases CNT Hazard Probe v1, a reusable hazard–memory analysis pipeline applied here to a synthetic accelerator-like system with beam-dump–style events. It is part of the ongoing development of Cognitive Nexus Theory (CNT), which hypothesizes that many complex systems share common “drift geometries” as they approach instability. In this experiment, a toy circular accelerator is simulated with: A noisy baseline signal with slow drift and low-frequency structure. “True” bursts that culminate in dump events (analogous to beam loss leading to a machine protection dump). “False” bursts that look similar but do not end in a dump. From a single scalar observable x(t)x(t)x(t) (a proxy for beam loss or machine stress), the CNT Hazard Probe constructs: A standardized field z(t)z(t)z(t) via global normalization of x(t)x(t)x(t). A hazard band defined by a quantile of ∣z∣|z|∣z∣ (here the 90th percentile), so that “hazard occupancy” is the fraction of samples in a window whose ∣z∣|z|∣z∣ lies in this band. Sliding windows of length Δ\DeltaΔ, each labeled according to whether an event occurs in the next horizon HHH samples (“event in [t, t+H]?”). For each window, a small feature vector is computed: mean_x — mean of the signal in the window std_x — local volatility slope — linear trend of x(t)x(t)x(t) over the window max_z — maximum standardized excursion in the window hazard_occupancy — fraction of points with ∣z∣|z|∣z∣ above the hazard-band threshold Using only these local features, the probe trains a low-complexity logistic regression model and evaluates: Time-respecting AUC Train on the early part of the time series, test on the later part (no temporal leakage). Event-holdout AUC Train on windows leading up to the first events and test on windows leading up to the final event, approximating a “new run / new year” scenario. Surrogate AUC distribution Shuffle x(t)x(t)x(t) while preserving event times, recompute features, and obtain a surrogate AUC distribution to estimate the chance-level baseline and the ΔAUC gap between real and surrogate series. The probe scans a grid of window sizes Δ\DeltaΔ and prediction horizons HHH. For each pair (Δ,H)(\Delta, H)(Δ,H), it records: Multi-feature AUC (auc_all) Single-feature baselines (auc_hazard_only, auc_std_only, auc_maxz_only) Surrogate mean and standard deviation ΔAUC = AUC_real − AUC_surrogate_mean Event-holdout AUC From this grid, a canonical hazard band is defined as the set of (Δ,H)(\Delta, H)(Δ,H) points where: auc_all ≥ 0.80 delta_auc_all ≥ 0.20 H ≤ 120 samples (to restrict to finite-range prediction) In the synthetic accelerator, this band falls roughly in: Window lengths: Δ≈10–120\Delta \approx 10–120Δ≈10–120 samples Horizons: H≈30–120H \approx 30–120H≈30–120 samples Within this region, the probe consistently achieves AUC ≳ 0.9 with ΔAUC ≳ 0.3–0.5 above surrogate baselines. At long horizons (H = 300), AUC values collapse toward chance, indicating a finite predictive range for pre-instability dynamics. Importantly, both the hazard-occupancy metric (quantile band on ∣z∣|z|∣z∣) and the max_z extreme-value metric contribute meaningful signal, while surrogate tests confirm that this structure is tied to temporal organization rather than static statistics. This record is intended as a template and reference: The included CSVs and plots document the full hazard-memory surface and canonical band for the synthetic accelerator. The same CNT Hazard Probe can be applied to real systems (e.g., financial indices, seismicity, power-grid frequency, accelerator beam-loss monitors) by providing a time series in the simple schema: time, x, event_flag where event_flag = 1 marks the onset of a “hazard” or “event” according to the user’s chosen definition. By running this probe on multiple domains and comparing their canonical bands, one can begin to test CNT’s cross-domain hazard geometry hypothesis: that diverse systems share similar pre-instability structure in the (Δ,H)(\Delta, H)(Δ,H) plane. Files in this record (for the “Notes” or “Description” lower section) You can add this as a sub-section in the description or in the “Notes” field: Key contents evidence/accelerator_multi_run/tables/ accelerator_synth_hard_hazard_memory_surface_multi_runs.csv accelerator_synth_hard_hazard_memory_surface_multi_runs_agg.csv accelerator_synth_hard_canonical_hazard_band_multi_runs.csv evidence/accelerator_multi_run/plots/ accelerator_synth_hard_AUC_surface_multi_runs.png Additional diagnostic plots for the multi-run accelerator experiments. evidence/hazard_probe_single/tables/ accelerator_synth_hard_hazard_memory_surface.csv accelerator_synth_hard_canonical_hazard_band.csv evidence/hazard_probe_single/plots/ accelerator_synth_hard_AUC_surface.png File names may differ slightly by timestamp, but all CSVs and PNGs in these directories are part of the same CNT Hazard Probe v1 experiment.